Anthropic flags a new kind of AI cyber threat: self‑replicating “weight worms”

Anthropic, the AI lab behind the Claude family of models, has released a new security report warning about a hypothetical but increasingly plausible class of malware it calls “weight‑based replicators.” Instead of spreading traditional executable files, these attacks would propagate the trained parameters (weights) of an AI model together with just enough runtime software to execute it on compromised devices.

According to Anthropic, such an AI “worm” could, in principle, copy model weights from machine to machine, gradually building a distributed network of powerful models that no longer depends on a centralized data center. The company stresses that today this is still impractical due to model size, bandwidth constraints and hardware costs. However, ongoing trends toward model miniaturization and more capable consumer‑grade GPUs or NPUs could make this type of attack far more realistic over the coming years.

The report argues that conventional malware defenses—like scanning for known binaries—would be poorly suited to detect weight‑based threats. Instead, organizations may need mechanisms to verify the provenance and integrity of AI models they run, along with behavioral monitoring tailored to model inference. Anthropic urges governments, cloud providers and enterprises to start treating model artifacts as sensitive assets in the software supply chain and to begin planning defenses before such techniques move from theory to practice.